Lexical Ambiguity 
and 
The Role of Knowledge Representation in Lexicon Design 
Branimir Boguraev 
Lexical Systems Group 
IBM T.J. Watson Research Center 
Yorktown Heights, New York 10598 
James Pustejovsky 
Computer Science Department 
Brandeis University 
Waltham, MA 02254 
Abstract 
The traditional framework \['or ambiguity res- 
olution employs only 'static' knowledge, ex- 
pressed generally as selectional restrictions or 
domain specific constraints, and makes uo use 
of any specific knowledge manipulation mech- 
anisms apart from the simple ability to match 
valences of structurally related words. In con- 
traust, this paper suggests how a theory of lex- 
ical semantics making use of a knowledge rep- 
resentation framework offers a richer, more ex- 
pressive vocabulary for lexical information. In 
particular, by performing specialized inference 
over the ways in which aspects of knowledge 
structures of words in context c~Ln be composed, 
mutually compatible and contextully relevant 
lexieal components of words and phrases are 
highlighted. In the view presented here, lexi- 
cal ambiguity resolution is an integral part of 
the same procedure that creates the semantic 
interpretation of a sentence itself. 
Keywords: lexical ambiguity, lexical semantics, compo- 
sitionality, lexical organization. 
1 Importing Knowledge 
Representation into the Lexicon 
Our thesis below is that a theory of lexical seman- 
tics making use of a knowledge representation (KR) 
framework offers a richer, more expressive vocabu- 
lary \['or lexical information. Ultimately the goal of 
this research is to explain the creative use of lan- 
guage and highlight the role of a constantly evolving 
lexicon, while obviating the current prevalent views 
of 'static' lexicon design. A side effect of adopting 
such a theory as the basis of semantic interpretation 
is that some classically difftcult problems in ambigu- 
ity --. in particular, lexical -- are resolved by viewing 
them from a different perspective. In an implemen- 
tation such as that proposed here, lexical ambiguity 
resolution is an integral part of the same procedure 
that produces the semantic interpretation of a sen- 
tence. 
There are several methodological motivations for 
wanting to import tools developed for computational 
representation and manipulation of knowledge into 
the si, udy of word meaning, or lexical semantics. In 
particular, we believe that the goals of computational 
linguistics are the same as those of linguistics: to pro- 
vide useful, testable and explanatory theories of the 
nature of language and its relation to human cogni- 
tion a.s a whole. It follows that computational lin- 
guistics is linguistics as it should now be done, and 
that the computational tools developed and available 
in the larger context of Artificial Intelligence should 
not be ignored by linguists, l"nrthermore, 'shifting' 
the application area of KR formalisms fi:om their tra- 
ditional domain (general world knowledge) to a level 
below words (lexical knowledge), allows us to ab- 
stract the notion of lexical meaning away from world 
knowledge, as well as from other semantic influences 
(e.g. discourse and pragmatic factors); such a process 
of abstraction is a crucial prerequisite of any theory 
of lexieal meaning. 
Bringing in KR tools provides us with several ben- 
efits, which are inst, rumental to enriching the seman- 
tics of lexical expressions. Firstly, it is now possible 
to systematically iucorpora.te world knowledge into 
the lexical entry, while still maintaining an aware- 
ness of the boundary between lexical and common 
sense knowledge. Secondly, it is also possible to rea- 
son over that knowledge, facilitating the construction 
of richer semantic interpretations. I"inally, having de- 
veloped a theory incorporating generalizations about 
the systematic patterning of words, we have a formal 
language for expressing these generalizations. The 
interplay of these capabilities results in a generative 
language for expressing the meanings of words, while 
providing a different way of capturing multiple word 
senses through richer composition. Together with a 
set of principles for lexical decomposition, whose cen- 
tral tenet is that semantic expressions for word mean- 
ing (in context) are constructed by a fixed number of 
generative devices (cf. Pustejovsky \[1989b\]), this lan- 
guage becomes a tool for expressing lexical knowl- 
edge, while not presupposing finite enumeration of 
word senses. 
Current dictionaries reflect, through their organi- 
zation, the traditional view of word senses; in particu- 
lar, they assume that the space of possible/allowable 
uses of a word is exhaustively carved out by an enu- 
merable set of senses \['or that word. Computational 
lexicons, to date, generally tend to follow this orga- 
36 I. 
nization. As a result, the natural language interl)re- 
tation tasks these lexica support acquire (or inherit) 
similar view to lexical ambiguity, which then necessi- 
tates a particular approach to disambiguation. Fur- 
thermore, dictionaries and lexicons currently are of 
a distinctly static nature: the division into separate 
word senses not only precludes permeability; it also 
fails to account for the creative use of words in novel 
contexts. In contrast, rather than taking a 'snapshot' 
of language at any moment of time and freezing it 
into lists of word sense specifications, the model of 
the lexicon proposed here does not preclude extend- 
ability: it is open-ended in nature and accounts for 
the novel, creative, uses of words in a variety of con- 
texts by positing procedures for generating semantic 
expressions for words on the basis of particular con- 
texts. 
In tbe remainder of this paper we will illustrate 
a particular theory of lexical semantics, following 
Pustejovsky \[forthcoming\] which promotes the notion 
of a generative lexicon. In particular, we brietly dis- 
cuss certain types of lexical ambiguity, demonstrate 
how traditional methods of mnbiguity resolution fail 
to scale up for these (and other) cases, and then out- 
line an approach to sernantie interpretation embody- 
ing richer methods of compositionality. 
As we also show below, the lexical model we pro> 
pose has the effect of greatly reducing the size of the 
lexicon. Moreover, it bears directly on issues of o> 
ganization and content of computational lexicons, as 
the model now embodies strong assumptions about 
the kinds of lexical aspects of words essential for nat- 
ural language processing. The generative theory of 
lexical semantics, then, imposes a strong focus on 
current efforts to extract lexical data from large on- 
line text resources (dictionaries and corpora): it not 
only offers a uniform representational framework for 
expressing the data extracted by the tools and meth- 
ods of computational lexicography (cf. Boguraev and 
Briscoe, \[1989\]), but also offers guidance on tl, e kinds 
of lexical data -- or distinctions in ttle lexical be- 
havior of words --- which should be sought in such 
resources (cf. Boguracv et al., \[1990\]). 
2 The Nature of Lexical Ambiguity 
One of the most pervasive phenomena in natural lan- 
guage, and one which every realistic language pro- 
ccssing application faces, is that of ambiguity. Con- 
sequently, resolution of lexical ambiguity becomes an 
essential task, without which deeper (or perhaps any) 
language understanding and interpretation is impos- 
sible. 
2.1 Inadequacy of Resolving Ambiguity by 
Enumeration 
As we pointed out earlier, all current computational 
lexicons lit into a particular processing framework 
for dealing with this problem. Assuming a partition.- 
ing of the space of possible uses of a word into word 
senses -- as postulated and defined by the entry for 
that word -- the problem becomes that of selecting, 
on the basis of various contextual factors (typically 
subsumed by, but not necessarily limited to, the no- 
tion of selectional restrictions), the word sense closest 
to the use of the word in tile given text.. Computa- 
tionally, this reduces to a search within a finite space 
of possibilities. 
Realistically, however, this approach fails on sev- 
eral accounts -- both in terms of what information 
is made available in a lexicon to drive the disam- 
biguation process and how an autonlated sense selec- 
tion procedure might make use of this information. 
Typically, external contextual factors alone are not 
sufficient for precise selection of a word sense; ad- 
ditionally, often the lexical entry does not provide 
enough reliable pointers to critically discriminate be- 
tween word senses. Secondly, the search process be- 
comes computationally expensive, if not in effect in- 
tractable, when it has to account for longer phrases 
made up of individually ambiguous words. Finally, 
the a~ssumption that an exhaustive listing can be as- 
signed to the different uses of a word lacks the ex- 
planatory power necessary for making generalizations 
and/or predictions about how words used in a novel 
way can be reconciled with their currently existing 
lexical definitions. 
To illustrate this last. point, below we present some 
examples of problematic nature for current ambiguity 
resolution frameworks. 
Creative Use of Words Consider the ambiguity 
and context-dependence of adjectives such as fast and 
slow, where tile meaning of the predicate varies de- 
pending on the head being modified. Typically, a 
lexicon requires an enumeration of di/ferent senses 
for such words, in order to account for the ambiguity 
illustrated below: 
h a fast car: Ambiguous: a car driven quickly / 
one that is inherently fast. 
2: a fast typist: the person performs the act of typ- 
ing quickly, 
3: a fast waltz: the motion of the dance is quick. 
4: a fast book: one that calJ be read in ashort time, 
5: a fast reader: one who reads quickly. 
These examples involve at least three distinct word 
senses for the word fasl: 
fast(l): to move quickly; 
fast(2) : to pertbrm some act quickly; 
:fast(3) : to do something that takes little time. 
(Note that in a real lexicon, word senses would be 
further annotated with selectional restrictions; these 
are omitted here for brevity.) Upon closer analysis, 
each occurrence of'fast abow., predicates in a slightly 
different way. In fact, any finite ennmeration of word 
senses will not account for creative applications of 
this adjective in the language. For example, fast in 
the phrase a fast motorway refers to the ability of 
vehicles on the motorway to sustain high speed. As 
a novel use of fast, we are clearly looking at a new 
sense that is not covered by the enumeration above. 
Permeability of Word Senses Part of our ar- 
gument for a different organization of the lexicon is 
based on a claim that the boundaries between the 
2 37 
word senses in the analysis of fast above are too rigid. 
Still, even if we assume that ennmeration is adequate 
as a descriptive mechanism, it is not always obvious 
how to select the correct word sense in any given con- 
text: conAder the systematic ambiguity of verbs like 
bake (discussed by Atkins el el., \[1988\]), which re- 
quire discrimination with respect to change-of-state 
versus create readings: 
6: John baked the potato. 
7: John baked the cake. 
The problem here is that there is too much overlap in 
the 'core' semantic components of the different read- 
ings; hence, it is not possible to guarantee correct 
word sense selection on the basis of selectional re- 
strictions alone. Furthermore, as language evolve~s, 
partial overlaps of core and peripheral components of 
different word meanings make the traditional notion 
of word ,~ense, as implemented in current dictionar- 
ies, inadequate (see Atkins \[1990\] for a critique of the 
tlat, linear enumeration-based organization of dictio- 
nary entries). The only feasible approach would be to 
employ considerably more refined distinctions in the 
semantic content of the complement than is conven- 
tionally provided by e.g. the mechanism of selectional 
restrictions. 
Difference in Syntactic Forms It is equally arbi- 
trary to create separate word senses for a lexical item 
just because it can participate in distinct syntactic 
realizations -- and yet this has been the only ap- 
proach open to computational lexicons which assume 
the ambiguity resolution framework outlined above. 
A striking example of this is provided by verbs such as 
believe and forget. Observe in (8--11) below that the 
syntactic realization of the complement determines 
both the factivity of the proposition in the comple- 
ment or whether an NP is interpreted as a concealed 
question (see, for example, Grimshaw \[1979\]). 
8: Mary forgot that she locked the door. 
9: Mary forgot to lock the door. 
10: Mary forgot the answer. 
11: Mary forgot her wallet. 
Sensitivity to factivity would affect, for instance, 
the interpretation by a question-answering system, 
when asked Did Mary lock the door? Since sentence 
(8) is factive and (9) is nonfactive, the answers should 
be Yes and No respectively. Such a distinction could 
be easily accounted for by simply positing separate 
word senses for each syntactic type, but this misses 
the obvious relatedness between the two instances of 
forget. If it were possible to make the use of forget 
in (8) and (9) sensitive to the syntactic type of its 
complement, then we could also explain the parallel 
cases in (10) and (11). This would allow us to have 
essentially one definition for forget which could, by 
suitable composition with the different complement 
types, ~eneratc all the allowable readings (cf. Puste- 
jovsky \[ 1989a\]). 1 
1 Note that such a treatment is different, in effect, from 
proposals to confiate more than one syntactic realization 
of the same word sense through the mechanism of type 
2.2 Towards a Dynamic Model of the 
Lexicon 
The major thrust of our analysis has attempted to 
show that the ambiguities shown above cannot be ad- 
equately handled by exhaustive enumeration of what 
are regarded as different word senses. It follows that 
the conventional computational framework for lexical 
ambiguity resolution, and in particular, the format 
for lexical entries in current computational lexicons, 
fails in at lea,st two respects. It is unable to describe 
all the 'senses' of a word through finite enumeration; 
and it is also unable to capture interesting general- 
izations between 'senses' of the same word. 
Such failures are partially due to limited (lexical) 
knowledge made available to natural language pro- 
cessing systems, as well as to an impoverished notion 
of (lexical) inference. Thus, the traditional frame- 
work for ambiguity resolution only employs 'static' 
knowledge, expressed as e.g. selectional restrictions, 
and no specific knowledge manipulation mechanisms 
apart from the simple ability to match valences of 
connected words. In contrast, we show below how 
a lexical entry can be assigned a richer knowledge 
structure and how, by performing specialized infer- 
encc over the ways in which aspects of knowledge 
structures of words in context can be composed, mu- 
tually compatible and relevant lexical components of 
words and phrases are highlighted. This process, li- 
censed by constraints operating through the inference 
mechanislns, in fact, results in generating a seman- 
tic interpretation of a phrase, resolving en route the 
ambiguity of lexical items at their source. 
3 Ambiguity and CompositionaHty 
The richer structure for the lexical entry proposed 
here takes to an extreme the established notions 
of predicate-argument structure, primitive decompo- 
s/tion and conceptual organization; these are then 
viewed as defining a space of possible contexts in 
which a word can be used. Rather than committing 
to an enumeration of a pre-determined number of dif- 
ferent word senses, a lexical entry for a word now 
encodes a range of deeper aspects of lexical mean- 
ing. Looking at a word in isolation, these meaning 
components simply denote the sernantic boundaries 
appropriate to its use. Viewing a word in the con- 
text of other words, mutually compatible aspects in 
the respective lexical decompositions become more 
prominent, thus forcing a specific interpretation of 
each individual word. It is important to realize that 
this is a generative process, which goes well beyond 
the simple matching of features. On the contrary, 
such a framework requires, in addition to a flexible 
notation for expressing semantic generalizations at 
the lexical level, a mechanism for composing these 
individual entries on the phrasal level. 
To get a better understanding of how the distinc- 
tions in lexical meaning manifest themselves, it is im- 
coercion -- see Briscoe et el. \[1990\] for a computationa\] 
approach, based on a suitably enriched lexical represen- 
tation and utilizing the notions of type coercion (Puste- 
jovsky, \[1989a\]) and quzdia structure (see below). 
38 3 
portant to study and detine the role that all lexical 
types play in contributing to the overall meaning of a 
phrase. Thi,~ is not just a methodological point: cru- 
cial to the processes of semantic interpretation which 
the lexicon is targeted for is the notion of composi- 
I.ionality, necessarily different from the more conven- 
tional pairing of e.g. verbs (as functions) and nouns 
(~ks arguments). As we indicated earlier, if the seman- 
tic load in the lexicon is entirely spread among the 
verb entries -- as many existing computational sys- 
tems assume --- differences like those exemplified in 
(6-7) and (8-11) can only be accounted for by treat- 
ing bake, forget, and so forth as polysemous verbs. 
lf, on the other hand, elaborate lexical meanings of 
e.g. verbs and adjectives could be made sensitive to 
componenLs of equally elaborate decompositions of 
e.g. nouns, the notion of spreading the semantic load 
evenly across the lexicon becomes the key organizing 
prirleiple in expressing the knowledge necessary for 
disambiguation. 
In order to be able to express the distinctions, at 
lexical level, required for analyzing the examples ill 
the last section, it is necessary to go beyond view- 
ing lexical decomposition as base.d only on a 1)rc- 
determiued set of primitives; rather, what is needed 
is the con bmction of being able to sl;ecify (e.g. by 
means of sets of predicates) different levels, or per- 
~;pectives, of lexical representation alid being able to 
compose (via a fixed numbe:" of generative devices) 
these predicates. A 'static' definition of a word now 
only provides its literal meaning; suitaMe compo- 
sitions of apl)ropriately 'highlighted' \[)rejections of 
(syntactically) related words, generate meanings in 
context. 
In such a way, many of the short coufings in 
particular those from the perspeetiw', of automatic 
language processing - of the more descriptive ap- 
proach inlmrent in exhaustive enunw.ration of word 
senses can be overcome. What makes this possible is 
the combination of two notions, both of them tbllow- 
ing from general principles of KIt theory. First, by 
positing a language R)r describing dif\[crent levels of 
word meanings, we are no longer confined to the con- 
;:~traints following from operating with a fixed inven- 
tory of primitives; moreover, we now also haw.' a way 
of incorporating in this language l,he set of rules gov- 
erning the generative processes. Secondly, through 
the very nature of these rules, we are assured that the 
~:~emantic representations ultimately associated with 
text (fragments) are going to be well-formed. 
3.1 Lew?ls of Lexical Meaning 
Pustejovsky \[t989b\] proposes several levels of lexi- 
cal representation. Following an analysis of a broad 
range of (traditionally ambiguous) constructions, and 
in particular the aspects of word meanings which ac- 
count for the ambiguities, he argues for four struc- 
tures that a theory of computational lexical seman- 
tics needs to capture. 
Argmnent Structure This defines the c.ouven- 
tional mapping fi'om a word to a function, and relates 
the syntactic realization of a word to tile number and 
type of arguments that are identified at the level of 
syntax and made use of at tile level of semantics (cf. 
Grimshaw \[1990\]). 
Event Structure This identifies the particular 
event type for a ve,'b or a phrase. There are essen- 
tially three components to this structure: the primi- 
tive event type --state (S), process (p) or transition 
(T); the focus of the event; and the rules for event 
composition. 
Qualia Structure This defines the essential at- 
tributes of an object associated with a lexical item. 
By positing separate components (see below) in what 
is, in essence, argument structure for nominals, nouns 
are elevated from the status of being passive argu- 
ments to active functions. 
Lexical Inheritance Structure This determines 
the way(s) in which a word is related to other con- 
cepts in tile lexicon. In addition to answering ques- 
tions concerning the organization of a (lexical) knowl- 
edge base, this level of word meaning makes it possi- 
ble to link lexieal knowledge with general world (com- 
mon sense) knowledge. 
Since the only level of lexical representation not 
extensively discussed in the literature is that of qualia 
structure, we briefly outline its components below. 
3.2 Qualia Structure 
The essence of tile proposal is that there is a sys- 
tem of relations that characterizes the semantics of' 
nominals, very much like the argument structure of 
a verb. Pustejovsky \[1989b\] calls this the Qualia 
Structure, inspired by Aristotle's theory of explana- 
tion and ideas from Moravesik \[1975\]. In effect, the 
qualia structure of a noun determines its meaning tus 
much as the list of arguments determines a verb's 
meaning. The elements that make up a qualia struc- 
ture include notions such as container, space, surface, 
figure, artifact, and so on. 2 
Briefly, the. Qualia Structure of a word specifies 
\['our tLspects of its meaning: 
o the relation between an object and its con- 
stituent parts; 
o that which distinguishes it within a larger do= 
main; 
o its purpose and fum;tion; 
o factors inw)lved iu its origin or "bringing it, 
about". 
These aspects of a word's meaning are called its 
Constitutive Role, I'brmal Role, 2blic Role, and 
Agentive Role, respectively, a The motivation tbr 
>I'hese components of an object's denotation have 
long been considered crucial for our commonsense under- 
standing of how things interact in the world. Cf. Hayes 
\[1979\], tlobbs et aL \[1987a\], and Croft \[1986\] for discus- 
sion of the,;e qualitative aspects of me,thing. 
agorae of these roles arc reminiscent of descriptors 
used by various comptttational researchers, such as Wilks 
\[1975\], llayes \[1979\], ~nd Ilol,bs et al. \[1987a\]. Within 
the theory outlined here, these roles determine a minimal 
semantic description of ~t word which has both semantic 
~tnd gra.mm~tieal consequences. 
4 39 
positing such characterizations of word meaning is 
that by enriching the semantic descriptions of nom- 
inal types, we will be able to "spread the seman- 
tic load" more evenly through the lexicon, while ac- 
counting for novel word senses arising in syntactic 
composition. 
3.3 Lexlcal Ambiguity Resolution 
Let us examine how this view is able to account 
for the ambiguities discussed in the previous section. 
Consider first the example with fast We can capture 
the general behavior of how such adjectives predicate 
by making reference to the richer internal structure 
for nominals suggested above. That is, we can view 
fast as always predicating of the :1bile role of a nom- 
inal. To illustrate this, consider the qualia structure 
for a noun such ~us car: 
car(*x*) 
\[Coast: 
\[Form: 
\[Telic: 
\[Agent: 
{body, engine .... }\] 
car- shape (* x* ) \] 
move(P,*x*), drive(P,y,*x*)\] 
artifact (*x*)\] 
Notice that the Tell( role specifies the purpose and 
function of the noun. In the phrase, a fast car, it 
is the relation specified there (seen as an event -- 
namely, a process, P) which is modified by tile ad- 
jective as being fast. Similarly, for the nouns typ- 
ist, waltz, book, and reader, it is their 'relic role that 
is interpreted as being fast (without going into de- 
tails, we note here that the Telic role of lypis! deter- 
mines the activity being performed, namely typing; 
similarly for waltz, its Telic role retZrs to (lancing). 
tlence, the interpretations of fast in the examples (1- 
5) above can all be derived from a siugle word sense, 
and there is no need for enumerating the different 
senses (cf. Pustejovsky \[forthcoming\]). The lexical 
semantics for this adjective will indicate that it acts 
as an event predicate, modifying the 'relic role of the 
noun, as illustrated in the minimal lexical semantic 
structure for fast below: 
fast(,x.) ~ (Tell(: AP3E\[fast(E) A f)(E, .x.)\]) 
Notice that, in addition to obviating the need for 
separate senses, we can generate the novel use of fast 
mentioned above in the phrase a fast motorway, since 
the Tell( role of mot.orway specifies its purpose, and 
it is this activity which is interpreted ~ fast: 
\[Tdic : travd(P, cars) A on(P, *x*)\]. 
The composition of the expression defining fast with 
the lexical aspect it specifies ms its 'target' -- the 
'relic role of its argument (motorway) -- results in 
an interpretation corresponding to a use of the word 
when referring to a road: one that allows for fast 
travel by cars. 
3.4 Implications for Natural Language 
Processing 
The framework proposed above is very attractive for 
NLP, for at lea~st two reasons. Firstly, it can be for- 
malized, and thus make the basis for a computational 
procedure for word interpretation in context. Sec- 
ondly, it does not require the notion of exhaustive 
enumeration of all the different ways ill which a word 
can behave, in particular in collocations with other 
words. Consequently, the fi:amework can naturally 
cope with the 'creative' use of language; that is, the 
open-ended nature of word combinations and their 
associated meanings. 
The method of fine-grained characterization of 
lexical entries, as proposed here, effectively allows 
us to conflate different word senses (in the tradi- 
tional meaning of this term) into a single meta-ent, ry, 
thereby offering great potential not only for system- 
atically encoding regularities of word behavior de- 
pendent on context, but also for greatly reducing 
the size of the lexicon. The theoretical claim here is 
that such a characterization constrains what a possi- 
ble word meaning can be, through the mechanism of 
logically well-formed semantic expressions. The ex- 
pressive power of a Kl~. formalism can then be viewed 
as simply a tool which gives substance to this claim. 
4 Knowledge Representation and 
Global Organization of Lexicon 
So far we have looked at the "classical" problem of 
ambiguity of words, manifested in the problem of how 
to select suitable word senses for a word in running 
text, according to some notion of context. As we 
pointed out just now, the solution outlined in the 
previous section, in addition to offering an alternative 
way of approachng the problem, has the important 
'side effect' on the size of the lexicon. 
In this section we address, at more depth, the ques- 
tion of how the techniq~es and methods of KR relate 
directly to the problem of lexical ambiguity resolu- 
tion, the information to bear on it, and the methods 
for solving it. The discussion is carried out in the con- 
text of an alternative manifestation of the ,~ame prob- 
lem, which we refer to as "hiddell" lexical ambiguity. 
We also use the shine context for presenting, intbr- 
really, some aspects of the lexical inheritance struc- 
ture as another level of lexical meaning. 4 
One of the implications of positing quails struc- 
tures is the necessity to have, superimposed on the 
lexicon, a realization of more than one lattice struc- 
ture. Earlier attempts at conceptual hierarchies faced 
this problem all the time: conceptual models typi- 
cally make heavy use of multiple inheritance: "book" 
is_a "literature", "book" is_a "object", "dictio- 
nary" 2s_a "object", "dictionary" is_a "reference", 
"car" is_a "vehicle .... car" is_a "artifact", and so 
4Introducing inheritance into the lexicon is not an en- 
tirely new idea. For example, Flickinger el al. \[1985\] dis- 
cuss the value of inheritance as a representational device 
for capturing generalizations across classes of lexical en- 
tries. A further argument for the usefulness of inheritance 
mechanisms is provided by Briscoe et hi. \[1990\], who 
show how a mechanism of lexieal inference can augment 
a text analysis systexn which performs syntactic analy- 
sis and semantic interpretation by making reference to 
detailed lexical decomposition of entries in the style of 
Pustejovsky \[1989@ 
40 5 
forth. Still, as descriptive as such relations may ap- 
pear, models like these suffer from a very limited no- 
tion of lexical structure. Thus, even though elaborate 
mechanisms have been proposed to control and limit 
the flow of information along the e.g. generalization/ 
specialization links, there has been no theory to ei- 
ther (a) explain how to assign structure to lexical 
items, or (b) specify lexical relations between lexical 
items in terms of links between only certain ~pects 
of their respective lexical structures. Pustejovsky's 
theory of lexical semantics \[1989b\], with its several 
distinct, levels of semantic description, and in par- 
ticular the qualia structure, are relevant to just this 
issue. 
On this view, a lexical item inherits infornmtion 
according the qualia structure it carries. In this way, 
the different senses for words can be rooted into suit- 
able lattices. To illustrate this point, consider the 
two is._a relations below, and the differences in what 
relations the objects enter into. 
play (is_a book) (dictionary is_a book) 
read ok no 
buy ok ok 
consult no ok 
begin ok(?) no 
Lexical inheritance theory, oll the other hand, 
posits a separate lattice per role in the qualia struc- 
ture. Briefly, inheritance through qualia amounts 
to the following relations for this exarnple: book 
is_form phys-object, book is_telic literature, dic- 
tionary is-~orm book, dictionary is_relic refer- 
ence, book is._agent literature, dictionary is.agent 
compiled-material, play is_agent literature, play 
is_relic performance. With the qualia roles differ- 
entiating the lattice structures, we can exclude the 
unwanted inferences listed above. 
5 Conclusion 
We have outlined a framework for lexical semantic 
research that we believe earl be uselifl for both corn- 
pvtational linguists and theoretical linguists alike. 
We argued against the view that word meanings are 
fixed and inflexible, where lexical ambiguity must 
be treated by multiple word entries in the lexicon. 
Rather, the lexicon can be seen ~ a generative sys- 
tem, where word senses are related by logical opera- 
tions defined by the well-formedness rules of the se- 
mantics. In this view, much of the lexical ambigu- 
ity of highly ambiguous lexical items is eliminated 
because the semantic load is spread more evenly 
throughout the lexicon to the other lexical categories; 
furthermore, the lexical knowledge we propose as nee- 
essary for ambiguity resolution is seen ms factored out 
at different levels of lexical representation. We looked 
at two of these levels, qualia structnre and lexieal in- 
heritance, as they turn out to be of particular rele- 
vance to the structuring of the semantic information 
carried by e.g. nouns and adjectives, and applying 
it, via composition, to the construction of semantic 
interpretation of complex expressions. The methods 
underlying the analysis of ambignous phrases and the 
construction of corresponding semantic expressions 
make extended use of Kt{. devices and techniques. 

References 

\[Atkins, 1990\] B. Atkins. "Building a Lexicon: Rec- 
onciling An|son\]orphic Sense Differentiations in 
Machine-Readable Dictionaries", presented at 
BBN Symposium: Natural Language in the 90's ..... 
Language and Actio~l in the World. 

\[Atkins el al., 1988\] B. Atkins, J. Kegl, 
and B. Levin. "Anatomy of a Verb l'\]ntry", In- 
ternational Journal of Lexicography, 1(l). 

\[Boguraev and Briscoe, 1989\] B. Boguraev 
and E. Briscoc (Eds). Comp~ltational Lezieograpl~y 
for Natural Lan(.l~mge Processi~g, IIarlow, Essex, 
Longman. 

\[Boguraev el al., 1990\] B. Boguraev, 
R. Byrd, J. Klawms, and M. Neff. "From Struc- 
tural Analysis of Lexical Resources to Semantics 
in a Lexical Knowledge Base", t~C 15427, IBM 
T.J.Watson llesearch Center. 

\[Briscoe el al., 1990\] E. Ilriscoe, A. Copestake, and 
B. Boguraev. "l,exical Semantics via Lexicology", 
ms. University of (lambridge Computer Labora- 
tory and IBM T.J.Watson Research Center. 

\[Croft, 1986\] W. Croft. Categories and Relations i7 ~, 
5'y~tax: 7'he Clause-Level OrgaT~ization of h~for- 
mation, Ph.D. I)iss., Stanford University, 1986. 

\[Flickinger et al., 1985\] D. Flickinger, C. Pollard, 
and T. Wasow. "St.ructure-Sharing in Lexical Rep- 
resentation", Proceedings or 23rd Ammal Meeting 
of the ACI.., Chicago, IL. 

\[Grimshaw, 1979\] J. Grimshaw. "Complement Selec- 
tion and the Lexicon", Li~g~tistic h~quiry. 

\[tlayes, 1979\] P.tlayes. "Naive Physics Manifesto", 
in Expert 5'ystems i7~ the Micro-Electronic Age, 
Donald Mitchie (Ed), Edinburgh University Press, 
Scotland. 

\[ltobbs et at., 1987a\] J. llobbs, W. Croft, T. Davies, 
D. Edwards, and K. Laws. "Commonsense Meta- 
physics aud l,exical Semantics", Computational 
Linguistics, 13(3-4\]. 

\[Levin and l{apoport, 1988\] 
B. Levin and T. l{apoport. "Lexical Subordina- 
tion", Proceedi'ags of CLS 24, 275-289. 

\[Moravcsik, 1975\] J.M. Moravcsik. "Aita as Genera- 
tive Factor in Aristotle's Plfilosophy", Dialogue. 

\[Pustejovsky, 1989a\] J. Pustejovsky. "Type Coercion 
and Selection", paper at West Coast Conference 
on Formal Linguistics, Vancouver, 1989. 

\[Pustejovsky, 1989b\] J. Pustejovsky. "Tile Genera- 
tive Lexicon", ms. Brandeis University. 

\[Pustejovsky, forthcoming\] J. Pustejovsky. The Gen- 
erative Lexicon: A 73eory of Comp~ttational Lexi- 
cal Sema~tics, MIT Press, Cambridge, MA. 

\[Wilks, 1975\] Y. Wilks. "An Intelligent Analyzer and 
Understander for li;nglish", CACM, 18, 264-274.
